An Explainable Attention-Guided Convolutional Neural Network Framework for Melanoma Detection and Classification
TL;DR
The proposed framework offered a valid, meaningful and clinically relevant solution for automated melanoma screening and revealed high sensitivity to the detection of malignant lesions, as 154 out of 167 images of melanomas were correctly identified.
Abstract
Melanoma is a more aggressive type of skin cancer and early detection is critical to good clinical outcomes. Interpretation is difficult, however, due to the fact that dermoscopic images of melanomas and benign lesions can present with similar visual features, such as uneven pigmentation, border variation and texture variation. In this study, a CNN with explainable attention guidance framework is designed to detect and classify melanoma skin lesion images from dermoscopic images. The dataset HAM10000 (10,015 images) was split into 1,113 melanoma and 8,902 non-melanoma images in a binary classification task. The proposed framework combines a preprocessing, data augmentation, class-balanced learning, channel–spatial attention-based feature refinement, and explainability via Grad-CAM. Attention mechanism was applied to improve feature learning to focus on lesions and class-balanced learning was applied to prevent the bias of non-melanoma samples. The model's performance on the independent test set resulted in an accuracy of 94.81%, precision of 91.76%, recall/sensitivity of 92.21%, specificity of 95.30%, F1 score of 91.98% and ROC-AUC of 0.972. The results revealed high sensitivity to the detection of malignant lesions, as 154 out of 167 images of melanomas were correctly identified.The confusion matrix indicated a high sensitivity towards melanomas (154/167 images were correctly identified). The proposed framework was evaluated by comparative analysis and it was seen that all the basic CNN, VGG16, ResNet50, DenseNet121, MobileNetV2, and EfficientNetB0 models were outperformed by the proposed model. The ablation study indicated that the contributions of attention-guided learning, data augmentation, and class balancing were each progressively in the direction of performance improvements. The results also revealed that the Grad-CAM visualization primarily highlighted areas of the image that were relevant to the lesions, including the presence of irregular borders, asymmetry, and pigment variations. As such, the proposed framework offered a valid, meaningful and clinically relevant solution for automated melanoma screening.